Inspiration

AI responses are only as reliable as the context they receive. Attachpad was inspired by the gap between simply uploading a file and actually controlling what an MCP-connected assistant can see, search, prioritize, and retrieve.

What It Does

Attachpad is a context control center for MCP clients. Users can:

  • Upload files or import public webpages and PDFs
  • Arrange context manually
  • Assign priority to documents
  • Connect their workspace through a session-scoped MCP URL

Once connected, MCP clients can:

  • List available files
  • Search indexed content
  • Read individual files
  • Retrieve focused context based on user queries

How We Built It

Attachpad is built with:

  • React Router (Framework Mode)
  • Tailwind CSS
  • shadcn/ui
  • Drizzle ORM
  • SQLite

The entire application was developed using Codex, with different variants of GPT-5.6 selected based on the task at hand, from architecture and backend implementation to frontend development, debugging, and refactoring. This AI-assisted workflow enabled rapid iteration while maintaining a cohesive codebase.

Files are stored using a browser-local session ID. Imported PDFs and webpages are normalized into Markdown for consistent retrieval. SQLite FTS5 powers local full-text search without requiring embeddings or external APIs. The MCP server exposes retrieval tools over HTTP.

Challenges We Ran Into

Some of the biggest challenges included:

  • Keeping browser sessions and MCP session identities synchronized
  • Configuring SQLite reliably with Drizzle/libSQL
  • Safely importing public URLs while handling redirects and blocking private-network access
  • Converting PDFs into high-quality, agent-readable Markdown while preserving the original documents

Accomplishments We're Proud Of

  • Clean dashboard with manual context ordering and priority management
  • Direct PDF-link importing with automatic Markdown extraction
  • Fully local retrieval powered by SQLite FTS5 (no embedding APIs required)
  • MCP interface supporting both full-context retrieval and focused search tools
  • Session-scoped workspaces without requiring user accounts

What We Learned

We learned that MCP integration is not just about exposing tools—it is about making context easy to curate, organize, and control.

We also learned that:

  • Session handoff between a browser and a desktop MCP client should be explicit and predictable.
  • Simple local search can be remarkably effective when ranking signals remain transparent to users.
  • AI-native software development can dramatically accelerate building and iterating on complex full-stack applications when the right model is matched to each development task.

What's Next for Attachpad

Our roadmap includes:

  • Richer document parsers
  • OCR support for scanned PDFs
  • Improved relevance ranking
  • Optional local embeddings
  • Automatic refresh for imported web sources
  • Collections and workspace sharing
  • A secure authenticated mode for multi-device synchronization

Built With

Share this project:

Updates